
A November 6 project report describes training small LoRA adapters for particular coding-agent harnesses, including pi, OpenCode and Hermes. The team says the agents use a shared model with an adapter chosen for the harness. Read the original announcement and project details.
Why it matters
The authors report fewer tool calls on held-out smoke-test tasks while maintaining or improving accuracy. Those are preliminary results from the project’s own evaluation, not a general speed guarantee. Its workflow also inspects tool calls using safeagents before they run.
What to know
Adapter-specific behavior and a tool-checking layer are interesting for local coding workflows, but neither proves an agent is secure. Check the code, permissions, models and evaluation setup before trying it; we have not run it.
Reported from linked primary sources. No model or product was independently tested.



